RSNA 2024 Lumbar Spine Degenerative Classification - 4th Place Solution

This is the implementation of the 4th place solution (yu4u's part) for RSNA 2024 Lumbar Spine Degenerative Classification at Kaggle. The overall solution is described in this discussion.

Requirements

• 24GB x 2 VRAM (trained on GeForce RTX 3090 x 2).

Preparation

• Download the competition dataset from here and put them in input directory.
• Install Docker/NVIDIA Container Toolkit.
• Build the Docker image and enter the Docker container:

  export UID=$(id -u)
  docker compose up -d
  docker compose exec dev /bin/bash

• Login to wandb:

  wandb login

Training Process

Run all processes

./run.sh

Resulting models will be saved in saved_models directory. Oof files and weights for stacking models (LightGBM and XGBoost) will be saved in the current directory.

Training Process (Step-by-Step)

Axial Depth Model

# Create axial dataset for depth prediction.
python 01_create_axial_dataset.py --dirname axial_dataset_128 --img_size 128

# Train axial depth model.
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=axial_depth_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=32 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=axial_dataset_128 data.fold_id=0
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=axial_depth_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=32 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=axial_dataset_128 data.fold_id=1
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=axial_depth_fold2 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=32 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=axial_dataset_128 data.fold_id=2
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=axial_depth_fold3 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=32 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=axial_dataset_128 data.fold_id=3
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=axial_depth_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=32 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=axial_dataset_128 data.fold_id=4

# Create axial dataset for test.
python 05_create_test_axial_dataset.py --mode val --target axial

# Test axial depth model.
python 03_axial_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=0 model.resume_path=saved_models/axial_depth2_fold0 test.tta=True
python 03_axial_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=1 model.resume_path=saved_models/axial_depth2_fold1 test.tta=True
python 03_axial_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=2 model.resume_path=saved_models/axial_depth2_fold2 test.tta=True
python 03_axial_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=3 model.resume_path=saved_models/axial_depth2_fold3 test.tta=True
python 03_axial_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=4 model.resume_path=saved_models/axial_depth2_fold4 test.tta=True

Axial Keypoint Model

# Create axial keypoint dataset.
python 21_create_axial_keypoint_dataset.py --img_size 512 --dirname axial_keypoint_dataset_512

# Train axial keypoint model.
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=16 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=keypoint_v3_fold0 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=axial_keypoint_dataset_512 trainer.deterministic=False data.fold_id=0
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=16 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=keypoint_v3_fold1 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=axial_keypoint_dataset_512 trainer.deterministic=False data.fold_id=1
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=16 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=keypoint_v3_fold2 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=axial_keypoint_dataset_512 trainer.deterministic=False data.fold_id=2
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=16 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=keypoint_v3_fold3 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=axial_keypoint_dataset_512 trainer.deterministic=False data.fold_id=3
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=16 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=keypoint_v3_fold4 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=axial_keypoint_dataset_512 trainer.deterministic=False data.fold_id=4

# Create axial keypoint dataset for test.
python 25_create_test_keypoint_dataset.py --mode val

# Test axial keypoint model.
python 23_axial_keypoint_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=0 model.resume_path=saved_models/keypoint_v3_fold0 test.tta=True
python 23_axial_keypoint_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=1 model.resume_path=saved_models/keypoint_v3_fold1 test.tta=True
python 23_axial_keypoint_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=2 model.resume_path=saved_models/keypoint_v3_fold2 test.tta=True
python 23_axial_keypoint_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=3 model.resume_path=saved_models/keypoint_v3_fold3 test.tta=True
python 23_axial_keypoint_test.py test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=4 model.resume_path=saved_models/keypoint_v3_fold4 test.tta=True

Axial Classification Model

# Create axial classification dataset from prediction.
python 14_create_axial_cls_dataset_from_pred_v2.py --channel_num 5 --img_size 224
python 17_add_axial_cls_dataset_from_gt.py  --channel_num 5 --img_size 224

# Train axial classification model.
# EfficientNetV2S
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_attn_hflip_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_attn_hflip_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_attn_hflip_fold2 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_attn_hflip_fold3 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_attn_hflip_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3

# Swin-T
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_swin_t_fold0 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_swin_t_fold1 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_swin_t_fold2 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_swin_t_fold3 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=ch5_swin_t_fold4 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k

Sagittal1 Depth Model

# Create sagittal1 dataset for depth prediction.
python 06_create_sagittal1_dataset.py

# Train sagittal1 depth model.
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=sagittal1_depth2_fold0 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=16 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=sagittal1_dataset data.fold_id=0
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=sagittal1_depth2_fold1 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=16 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=sagittal1_dataset data.fold_id=1
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=sagittal1_depth2_fold2 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=16 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=sagittal1_dataset data.fold_id=2
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=sagittal1_depth2_fold3 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=16 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=sagittal1_dataset data.fold_id=3
python 02_axial_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=256 wandb.name=sagittal1_depth2_fold4 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 trainer.deterministic=False task.img_size=128 task.depth=16 model.ema=True model.ema_decay=0.9995 trainer.check_val_every_n_epoch=8 task.in_channels=3 task.layer_num=1 task.dirname=sagittal1_dataset data.fold_id=4

# Create sagittal1 dataset for test.
python 05_create_test_axial_dataset.py --mode val --target sagittal1

# Test sagittal1 depth model.
python 03_axial_test.py test.target=sagittal1 test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=0 model.resume_path=saved_models/sagittal1_depth2_fold0
python 03_axial_test.py test.target=sagittal1 test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=1 model.resume_path=saved_models/sagittal1_depth2_fold1
python 03_axial_test.py test.target=sagittal1 test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=2 model.resume_path=saved_models/sagittal1_depth2_fold2
python 03_axial_test.py test.target=sagittal1 test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=3 model.resume_path=saved_models/sagittal1_depth2_fold3
python 03_axial_test.py test.target=sagittal1 test.mode=val trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 task.img_size=128 task.in_channels=3 task.layer_num=1 data.fold_id=4 model.resume_path=saved_models/sagittal1_depth2_fold4

Sagittal1 Keypoint Model

# Create sagittal1 keypoint dataset.
python 28_create_sagittal1_keypoint_dataset_from_improved.py

# Train sagittal1 keypoint model.
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal1_keypoint_v2_bs8_lr2e-3_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_keypoint_dataset trainer.deterministic=False data.fold_id=0
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal1_keypoint_v2_bs8_lr2e-3_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_keypoint_dataset trainer.deterministic=False data.fold_id=1
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal1_keypoint_v2_bs8_fold2 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_keypoint_dataset trainer.deterministic=False data.fold_id=2
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal1_keypoint_v2_bs8_fold3 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_keypoint_dataset trainer.deterministic=False data.fold_id=3
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal1_keypoint_v2_bs8_lr2e-3_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_keypoint_dataset trainer.deterministic=False data.fold_id=4

# Create sagittal1 keypoint dataset for test.
python 27_create_test_sagittal1_keypoint_dataset.py

# Test sagittal1 keypoint model.
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal1 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=0 model.resume_path=saved_models/sagittal1_keypoint_v2_bs8_lr2e-3_fold0
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal1 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=1 model.resume_path=saved_models/sagittal1_keypoint_v2_bs8_lr2e-3_fold1
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal1 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=2 model.resume_path=saved_models/sagittal1_keypoint_v2_bs8_fold2
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal1 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=3 model.resume_path=saved_models/sagittal1_keypoint_v2_bs8_fold3
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal1 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=4 model.resume_path=saved_models/sagittal1_keypoint_v2_bs8_lr2e-3_fold4

Sagittal1 Classification Model

# Create sagittal1 classification dataset from prediction.
python 31_create_sagittal1_cls_dataset_from_pred_v2.py --mode val

# Train sagittal1 classification model.
# EfficientNetV2S
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_base_ema0.995_norm_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_base_ema0.995_norm_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_base_ema0.995_norm_fold2 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_base_ema0.995_norm_fold3 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_base_ema0.995_norm_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3

# Swin-T
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_swint_fold0 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_swint_fold1 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_swint_fold2 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_swint_fold3 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s1_swint_fold4 opt.lr=2e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal1_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k

Sagittal2 Keypoint Model

# Create sagittal2 keypoint dataset.
python 41_create_sagittal2_keypoint_dataset_from_improved.py

# Train sagittal2 keypoint model.
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal2_keypoint_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_keypoint_dataset trainer.deterministic=False data.fold_id=0
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal2_keypoint_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_keypoint_dataset trainer.deterministic=False data.fold_id=1
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal2_keypoint_fold2 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_keypoint_dataset trainer.deterministic=False data.fold_id=2
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal2_keypoint_fold3 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_keypoint_dataset trainer.deterministic=False data.fold_id=3
python 22_axial_keypoint_train.py --monitor val_keypoint_error trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=8 trainer.precision=16 data.num_workers=12 trainer.max_epochs=32 wandb.name=sagittal2_keypoint_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=1e-2 scheduler.min_lr=0 task.img_size=512 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_keypoint_dataset trainer.deterministic=False data.fold_id=4

# Create sagittal2 keypoint dataset for test.
python 42_create_test_sagittal2_keypoint_dataset.py --mode val

# Test sagittal2 keypoint model.
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal2 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=0 model.resume_path=saved_models/sagittal2_keypoint_fold0
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal2 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=1 model.resume_path=saved_models/sagittal2_keypoint_lr1e-3_fold1
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal2 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=2 model.resume_path=saved_models/sagittal2_keypoint_fold2
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal2 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=3 model.resume_path=saved_models/sagittal2_keypoint_fold3
python 23_axial_keypoint_test.py test.mode=val test.target=sagittal2 trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=32 task.img_size=512 data.fold_id=4 model.resume_path=saved_models/sagittal2_keypoint_fold4

Sagittal2 Classification Model

# Create sagittal2 classification dataset from prediction.
python 43_create_sagittal2_cls_dataset_from_pred_v2.py --channel_num 9

# Train sagittal2 classification model.
# EfficientNetV2S
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_base_ch9_lr1e-3_fold0 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_base_ch9_lr1e-3_fold1 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_base_ch9_lr1e-3_fold2 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_base_ch9_lr1e-3_fold3 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_base_ch9_lr1e-3_fold4 opt.lr=1e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3

# Swin-T
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_swint_ch9_fold0 opt.lr=1e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_swint_ch9_fold1 opt.lr=1e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_swint_ch9_fold2 opt.lr=1e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_swint_ch9_fold3 opt.lr=1e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_swint_ch9_fold4 opt.lr=1e-4 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.995 task.dirname=sagittal2_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k

# Axial
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_axial_v2s_fold0 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=0 model.in_channels=3 task.train_target=sagittal2
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_axial_v2s_fold1 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=1 model.in_channels=3 task.train_target=sagittal2
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_axial_v2s_fold2 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=2 model.in_channels=3 task.train_target=sagittal2
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_axial_v2s_fold3 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=3 model.in_channels=3 task.train_target=sagittal2
python 12_axial_cls_train.py trainer.accelerator=gpu trainer.devices=-1 trainer.strategy=ddp_find_unused_parameters_true data.batch_size=32 trainer.precision=16 data.num_workers=12 trainer.max_epochs=16 wandb.name=s2_axial_v2s_fold4 opt.lr=2e-3 opt.opt=AdamW opt.weight_decay=0 scheduler.min_lr=0 task.img_size=224 model.ema=True model.ema_decay=0.99 task.dirname=axial_cls_all_dataset loss.mixup=0.3 model.arch=2.5d data.fold_id=4 model.in_channels=3 task.train_target=sagittal2

Create Out-of-Fold Predictions

# Axial
# EfficientNetV2S
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=0 model.resume_path=saved_models/ch5_attn_hflip_fold0 test.tta=True
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=1 model.resume_path=saved_models/ch5_attn_hflip_fold1 test.tta=True
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=2 model.resume_path=saved_models/ch5_attn_hflip_fold2 test.tta=True
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=3 model.resume_path=saved_models/ch5_attn_hflip_fold3 test.tta=True
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=4 model.resume_path=saved_models/ch5_attn_hflip_fold4 test.tta=True

# Swin-T
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=0 model.resume_path=saved_models/ch5_swin_t_fold0 test.tta=True model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=1 model.resume_path=saved_models/ch5_swin_t_fold1 test.tta=True model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=2 model.resume_path=saved_models/ch5_swin_t_fold2 test.tta=True model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=3 model.resume_path=saved_models/ch5_swin_t_fold3 test.tta=True model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val data.fold_id=4 model.resume_path=saved_models/ch5_swin_t_fold4 test.tta=True model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k

# Sagittal1
# EfficientNetV2S
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=0 model.resume_path=saved_models/s1_base_ema0.995_norm_fold0
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=1 model.resume_path=saved_models/s1_base_ema0.995_norm_fold1
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=2 model.resume_path=saved_models/s1_base_ema0.995_norm_fold2
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=3 model.resume_path=saved_models/s1_base_ema0.995_norm_fold3
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=4 model.resume_path=saved_models/s1_base_ema0.995_norm_fold4

# Swin-T
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=0 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s1_swint_fold0
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=1 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s1_swint_fold1
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=2 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s1_swint_fold2
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s1_swint_fold3
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal1_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal1 data.fold_id=4 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s1_swint_fold4

# Sagittal2
# EfficientNetV2S
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=0 model.resume_path=saved_models/s2_base_ch9_fold0
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=1 model.resume_path=saved_models/s2_base_ch9_fold1
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=2 model.resume_path=saved_models/s2_base_ch9_fold2
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=3 model.resume_path=saved_models/s2_base_ch9_fold3
python 18_axial_cls_test.py trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=4 model.resume_path=saved_models/s2_base_ch9_fold4

# Swin-T
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=0 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_swint_ch9_fold0
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=1 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_swint_ch9_fold1
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=2 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_swint_ch9_fold2
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_swint_ch9_fold3
python 18_axial_cls_test.py --output_suffix swint trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=sagittal2_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 data.fold_id=4 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_swint_ch9_fold4

# Axial
python 18_axial_cls_test.py --output_suffix axial trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 task.train_target=sagittal2 data.fold_id=0 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_axial_fold0
python 18_axial_cls_test.py --output_suffix axial trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 task.train_target=sagittal2 data.fold_id=1 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_axial_fold1
python 18_axial_cls_test.py --output_suffix axial trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 task.train_target=sagittal2 data.fold_id=2 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_axial_fold2
python 18_axial_cls_test.py --output_suffix axial trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 task.train_target=sagittal2 data.fold_id=3 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_axial_fold3
python 18_axial_cls_test.py --output_suffix axial trainer.accelerator=gpu trainer.devices=[0] data.num_workers=10 data.batch_size=64 task.dirname=axial_cls_all_dataset task.img_size=224 model.arch=2.5d model.in_channels=3 test.mode=val test.target=sagittal2 task.train_target=sagittal2 data.fold_id=4 model.backbone=swin_tiny_patch4_window7_224.ms_in22k_ft_in1k model.resume_path=saved_models/s2_axial_fold4

# Convert predictions to submission format
python 51_create_submission.py

Train Stacking Models (LightGBM and XGBoost)

# Create stacking dataset
python 74_create_sep_stacking_dataset.py --mode val

# Train stacking models
python 85_train_lightgbm.py
python 86_train_xgboost.py